Beyond Performance Tools: Measuring and Modeling Productivity in HPC

Abstract

Emerging challenges to productivity are not well covered by traditional methods for evaluating HPC programs, systems, and practices. The common measure of merit widely put forward in High-Performance Computing (HPC), high computational performance as measured in floating-point operations per second (FLOPs), does not account for many bottlenecks in real HPC workflow that increase time to solution which are unaffected by performance changes. In this paper we discuss these bottlenecks, show an approach to analyzing productivity based on measurement and modeling of HPC workflow, and present plans for measurement and experimentation tools to study and improve productivity in HPC projects with large computational and data requirements.